How to Tell if an Image Is AI-Generated

There is no single test that tells you whether an image is AI-generated. What works is collecting several independent signals: what the file says about itself (metadata and watermarks), what the pixels look like (to a detector and to your own eyes), and where the image came from. Each signal can fail on its own. You can be confident only when several of them point the same way.

The short version: checks in order

1. Get the best copy. Ask for the original file, not a screenshot or a repost. Screenshots and messengers strip most of the evidence.
2. Read the provenance metadata. A signed Content Credentials (C2PA) manifest or an IPTC label is the strongest evidence a file can carry. Open it in the C2PA checker.
3. Check for an invisible watermark. If a Google model could have made the image, check SynthID in the Gemini app — the SynthID guide explains how.
4. Run a pixel detector such as the AI image detector when the file has no metadata.
5. Look for visual tells — hands, text, reflections, backgrounds.
6. Search for the image and check the context: who posted it first, and whether other photos of the same scene exist.

Visual tells, and why they are getting weaker

Generators used to give themselves away with obvious mistakes. Many of these still appear, especially in busy scenes:

• Hands and fingers: extra or merged fingers, joints bending the wrong way, hands that fade into objects.
• Text: shop signs, labels, number plates and book covers with letters that almost form words.
• Reflections and shadows: a mirror that shows a different scene, shadows falling in two directions, light with no visible source.
• Backgrounds: objects melting into each other, repeated faces in a crowd, windows and railings that do not line up.
• Paired details: earrings, glasses, buttons and collars that do not match left to right.
• Overall look: skin that is too smooth, and lighting that feels like a studio even outdoors.

Provenance metadata: C2PA, IPTC and generator parameters

Many generators and editors now label their output inside the file. There are three common forms:

• C2PA / Content Credentials. A signed manifest that records who created the file and what was done to it. OpenAI adds it to images from ChatGPT, Adobe Firefly attaches it to generated images, and Google adds it to some of its generated images. A generated image usually has a c2pa.created action with the source type trainedAlgorithmicMedia.
• IPTC digital source type in XMP metadata. The same vocabulary without a signature: some generators and editors write it on export.
• Generator parameters. Local tools such as AUTOMATIC1111 and ComfyUI store the prompt, seed and settings in PNG text chunks.

A C2PA signature on its own is not an AI sign: recent cameras and phones also sign ordinary photos. The deciding field is the source type. The C2PA checker shows the manifest, the signer and the recorded actions, but it does not verify the signature cryptographically — for that, use the official verify service at contentcredentials.org. To see every metadata block at once (EXIF, XMP, IPTC, PNG text), drop the file into the EXIF viewer and remover; the image watermark detector additionally searches for generator names and hidden-data anomalies.

Remember the asymmetry: metadata that says "AI" is a strong signal, but missing metadata means nothing. Screenshots, most messengers, many social networks and image optimizers drop these blocks, and anyone can delete them on purpose.

Invisible and visible watermarks

SynthID is Google DeepMind's watermark for content made with Google's AI. It is woven into the pixels at generation time rather than stored in metadata, and it is designed to survive compression, resizing and screenshots reasonably well. Only Google's detector can read it. The most accessible way to check is to upload the image to the Gemini app and ask whether it was made with Google AI. Keep the limits in mind: the check only covers Google tools and partners that use SynthID, and "not found" does not mean "made by a human".

Meta says it marks images from Meta AI with a visible marker, an invisible watermark and IPTC metadata, and its apps show an "AI info" label. That label is part of the app interface, not the file: a picture saved from a post usually comes back without it.

Visible marks such as the sparkle in the bottom-right corner of images from the Gemini app are the weakest signal. They are not added in every product or plan, a simple crop removes them, and anyone can paint one onto a real photo. Our Gemini watermark remover exists for people who need a clean visual on their own generated images, for a banner or a slide. It does not touch SynthID, so it does not hide where an image came from — and presenting a generated image as a real photo breaks the rules of Google and most platforms.

AI image detectors: useful, but probabilistic

A detector is a classifier trained on real and generated images. It picks up statistical traces of generation such as noise patterns and upscaling artefacts, and returns a probability, not a fact. Typical failure modes:

• False positives: real photos with heavy processing — HDR, night mode, AI noise reduction, beauty filters, upscalers.
• False negatives: partial edits (AI only replaced the sky), small or recompressed images, screenshots, and generators released after the detector was trained.
• Disagreement between services: each model has its own training data and threshold, so 12% in one tool and 90% in another is not unusual. Compare conclusions, not numbers.

Our AI image detector combines both layers: it reads the metadata and runs a classification model in your browser. Treat its score as one opinion among the signals above.

Reverse image search and context

Often the fastest answer is not inside the image at all. Run a reverse search in Google Lens or TinEye and look for the earliest and largest copy. Google's "About this image" panel can also show AI labels taken from metadata, where available. Then ask the journalist's questions: who published it first, is there an author with a history, are there other photos or videos of the same event from different angles, and do the weather, signs and clothing match the claimed place and date?

How to read the combined result